There's a whole other category here that isn't "sign up for another chatbot."
Open-weight models (plain English)
Some organisations publish the weights of a model — the numbers that define how it behaves — so anyone can download and run it themselves. The names you'll hear:
- Meta → Llama
- Mistral
- Qwen (Alibaba)
- Plenty more, via hubs like Hugging Face
And mind the easy mix-up: open-weight is not the same as OpenAI the company. OpenAI sells ChatGPT (rented). Meta publishes Llama (you run it yourself, with the right tool).
How people actually run it
On a normal laptop or PC, Ollama is a common starter app — one install, then you pull a model down by name and chat in the terminal, or wire it into an editor later.
This lesson doesn't install anything. We're mapping the rented cloud world first. The hands-on install — Ollama on your own machine — comes later, once you've compared the cloud tools and built solid privacy habits.
Local vs cloud — the honest comparison
| Rented cloud (ChatGPT, etc.) | Open-weight on your PC | |
|---|---|---|
| Privacy | Data on their servers | Stays on your machine (if you keep it local) |
| Ongoing cost | Subscription / API per message | Electricity + hardware you already own |
| Quality | Frontier models on hard tasks | Depends on model size; smaller ones trade quality for speed |
| Offline | Needs internet | Works offline once downloaded |
| Rules | They change features and limits | You control the versions (more work for you) |
Local isn't automatically "better" — that's the honest bit. What it is, is yours. I still reach for cloud sometimes. But I stopped paying for five different clouds doing the same job the moment I understood this split.
Hardware honesty
A rough guide for starter local chat (not training your own models):
- 8 GB RAM — small models only; can feel sluggish.
- 16 GB RAM — comfortable for the common 7B–8B class models.
- A dedicated GPU — faster, but absolutely not required to learn.
When we reach the install, we pick a sensible starter model for your machine — no gatekeeping nonsense about needing a £2k PC.
Continue — five minutes to actually feel prompting work, then a quick recap before the quiz.